Bound and monitor AI extraction workflows

[kelvinkipruto]

Parent

#572

What to build

Make AI extraction runs bounded, non-overlapping, traceable, and operationally visible. Every extraction can be matched to its downstream upload, duplicate cron runs cannot process the same work, and provider time/cost exposure is capped per call and document.

Acceptance criteria

• Every extraction row has a required system-generated unique ID, with existing records safely backfilled.
• The extraction workflow has an atomic queued/running guard preventing overlapping runs over the same work.
• AI calls enforce per-call deadlines and each document enforces total duration, character, chunk, token, retry, and cost ceilings.
• Rejected or truncated documents enter an explicit operator-reviewable state.
• Retries use bounded exponential backoff and respect provider rate-limit guidance.
• Metrics and alerts cover calls, tokens, duration, retries, costs, workflow pileups, and terminal failures.
• Tests cover missing IDs, duplicate scheduling, provider timeout, retry exhaustion, and oversized documents.

Blocked by

None - can start immediately.